But even unintentionally serving CSAM is unique in that it can get you in legal hot water reeeal quick. It’s why web services are so strict about filtering.
I just can’t wrap my head around this. Maybe I’m misreading what Grok did, or maybe the story isn’t quite right, but if it is, the company should be getting raked over hot coals by the feds right about now.
I agree with this.
That lines up on the technical side, too, with how LLM “intelligence” is plateauing, cost of cheaper models is decreasing, where open weights are going and such. Q3 2027 seems about when an OpenAI coding subscription makes no sense.
That being said, I’d be wary of the “Facebook effect.”
Once a service gains a huge foothold, it can deteriorate for a long time without going away. Especially with regulatory capture. And for many smartphone users, OpenAI is the only AI they know.
Horrors of this aside… is Grok using a keyword filter for CSAM detection?
I cannot even begin to describe how technically negligent that is.
Image based CSAM detectors have been around for ~2 decades, and any image host with an ounce of self preservation uses them. I’m pretty use some Lemmy instances use them. Twitter must have already been using them, even before Musk bought them, because it would be legal suicidal not to.
That they wouldn’t implement a strong image-based checker in Grok is mind-boggling.
Grok has as much guardrails against CSAM as any other Ai. This is money grab lawsuit and nothing more.
Apparently not.
CSAM detectors and image categorizers are very, very good. There’s no way this guy could have made so many images without tripping them, if the guardrails were actually appropriate.
Instead, it appears Grok used a simple keyword detector… that’s not how CSAM detection is supposed to work.
That is bullshit as fuck. At his point any moron with decent GPU can make images with no guardrails at all.
There’s still a technical barrier.
I know it seems trivial, but simply offering it and advertising the service on X opens it to orders of magnitude more people than self-hosted models.
Maybe, just maybe, engagement ranked social media wasn’t a good idea?
Well, because it was originally sold as a DRM free game.
handles this just fine
But I would dispute this. I’m working out a modlist right now, and I’d be pulling my hair out if I had to do it in game, or manually with the XML file + mod IDs.
It’s annoying, I hate it. They’re useless bloat.
But it’s never risen to the level of “I literally will not buy/play the game over this.”
I don’t get people who say this at all. You buy this cool game you love, you’ll spend hours playing it, but you won’t enter a fake email and wait another 30 seconds at launch to play it?
Even at the peak of Origin’s woes, I never got close to that.
Rimworld has no native mod manager, and TBH editing it in game is slow and clunky. You can edit the XML file, but this is true of many games.
Of course one can use RimSort, but that’s a 3rd party fan project. It’s not really fair to expect every game to have a good one.
Nah, he’s aware of what he’s doing.
He was just “engagementmaxxing at all costs” before it was en vogue.
Well, that’s because Zuckerberg felt FOMO and fired the Llama team over one bad experiment he probably pressured them into in the first place (as Llama 4 seemed like a “quick and dirty” attempt to copy Deepseek to me).
He hired literal narcissists in their place, who gave him big promises they couldn’t possibly keep, given their background.
Meta was literally at the center of open weights/open source ML land, which is where all this will eventually settle. Between that and PyTorch, they could have been at the center of the universe.
But they aren’t, because Zuck is so unbelievably insecure.
I’m not one to blame individuals for anything; systemic failure is complex. But he actually did this to himself.
And what’s bizarre is, somehow, it will work out for him.
I’ve been criticizing Facebook since I was in high school and it was all the rage. But here we are. Zuckerberg is more “revered” than ever.
That’s the suspicion.
To be more accurate, I’d argue the data isn’t yet strong enough to support the headline.
Well that’s often counterproductive, too. Even the largest models degrade if you fill their context with stuff they don’t need, or push them outside the patterns they were finetuned for.
Not that RAG isn’t really important.
By “these guys” you mean a separate team setting up an agentic workflow just for your workplace?
I’ve finetuned some models for a job, not just LLMs, and I can all but guarantee they aren’t finetuning some self hosted coding model. The APIs they use do not “learn” from mistakes, so if that’s what management is thinking, they are misinformed.
And interesting things like constrained output are often not even supported by APIs.
Modern LLM APIs and wrappers come with really good tool harnesses for coding, too. Like, good enough you don’t want to screw with it and pollute the context structure the original trainers tuned it for.
What’s more… not only do models change, but single model’s capabilities change under load, for more speed during peak usage times. So there is no guarantee some configuration they find that “works better” will work better later. See:
What I’m saying is… I wonder how long this mess last?
Because if you already have GH PR integration, most of what an agentic team can do is mess things up.
We have to do this so we can improve the AI process.
…To be clear, does this manager think y’all are “improving” the model with feedback? Like it will learn or train off your comments?
Or what? Are they just talking about the intermediate steps to the LLM API?
I’m just trying to wrap my head around this.
No manager is “gathering training data.” Virtually no company knows how to finetune anything, they just use Claude API.
If you’re interested in further reading, ISW sometimes covers this figures in their Ukraine updates.
You might be particularly interested in the nearly successful Wagner rebellion, where a Russian mercenary company drove a convoy to Moscow and almost made it, largely unopposed by the Russian military. Some joined him:
Putin was always paranoid, but this is the reason for the extra recent paranoia, FYI.
Interesting.
Yes that was my suspicion.
I think this would be a more solid trend if the 16-24 bracket ages, and ends up less intelligent when they’re 25-34.
Well, Putin is actually a “moderate” of sorts in Russian politics. There are some scary, more nationalist and extremist, figures he keeps in line that would be happy to fill the power vacuum.
It’s not even predicable anymore, as Putin’s already killed the most “obvious” popular ones, like Prigozhin.
That’s what I meant.
Maybe they could kill Putin, without him becoming a martyr. But Ukraine doesn’t know what comes after Putin.
In a country with so many nukes, that could be a literally catastrophic gamble.
I wonder if Ukraine got the idea from Iran?
I don’t mean that as an accusation, but… well, Iran basically has the world economy’s throat. And I know Ukraine’s drones just got the capability, but Iran had to be a fantastic reminder of how fragile an oil economy is.
Well, I bought 1 year of GLM’s coding plan for basically nothing ($50?) last year.
…I now see its many hundreds of dollars, and its API is so busy it gets throttled.
Seems like devs are catching on.
FYI, for anyone interested in this stuff, I would suggest 1 year of Xiaomi’s coding subscription.
MiMo 2.5 is fantastic, beyond what benchmarks show. It’s a genuinely useful general purpose model with better prose/world knowledge than others. It’s quite uncensored. It’s not just some deep fried agent, and it’s dirt cheap right now because no one has realized it yet (kind of like GLM 6 months ago).
And before you ask, I use local models too. I sometimes run a custom IQ3_KT of MiMo 2.5, in fact, which is how I figured out it’s good. self hosting is the way. Gestures at the forum we’re on
But for stuff where privacy isn’t so critical, it’s still nice to have a year of an unquantized, fast API of a huge model for like $50.